<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Swetabh Tripathy</title>
    <description>The latest articles on DEV Community by Swetabh Tripathy (@tripathyswetabh).</description>
    <link>https://dev.to/tripathyswetabh</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4162113%2Fd69767f8-8e6c-4bc5-a772-24bddee96f7c.png</url>
      <title>DEV Community: Swetabh Tripathy</title>
      <link>https://dev.to/tripathyswetabh</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/tripathyswetabh"/>
    <language>en</language>
    <item>
      <title>Pocket Ledger: A Local Receipt Sorter for Lazy Bookkeeping</title>
      <dc:creator>Swetabh Tripathy</dc:creator>
      <pubDate>Tue, 20 Oct 2026 11:41:19 +0000</pubDate>
      <link>https://dev.to/tripathyswetabh/pocket-ledger-a-local-receipt-sorter-for-lazy-bookkeeping-7gb</link>
      <guid>https://dev.to/tripathyswetabh/pocket-ledger-a-local-receipt-sorter-for-lazy-bookkeeping-7gb</guid>
      <description>&lt;h1&gt;
  
  
  Pocket Ledger: A Local Receipt Sorter for Lazy Bookkeeping
&lt;/h1&gt;

&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19"&gt;Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;Bookkeeping is one of those tasks that seems small until a month's worth of receipts piles up.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Pocket Ledger&lt;/strong&gt; is a local-first receipt parser that turns messy receipt text into a structured, reviewable CSV. It's designed for freelancers, consultants, designers, and developers who want to spend less time manually entering expenses and more time doing the work they actually enjoy.&lt;/p&gt;

&lt;p&gt;Users can paste receipt text or upload text files. Pocket Ledger extracts:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Vendor&lt;/strong&gt; — who received the payment.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Date&lt;/strong&gt; — when the transaction occurred.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Amount&lt;/strong&gt; — the expense value.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Expense category&lt;/strong&gt; — a suggested bookkeeping category.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Review flags&lt;/strong&gt; — fields that need human attention.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The idea is simple: automate the repetitive parts, surface uncertainty, and let the user make the final call.&lt;/p&gt;

&lt;p&gt;Rather than pretending every receipt can be parsed perfectly, Pocket Ledger makes uncertain records visible so they can be reviewed before export.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;You can run Pocket Ledger locally:&lt;/p&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;
bash
git clone https://github.com/Tswetabh/Husltr.git
cd Husltr

pip install -r requirements.txt
streamlit run app.py
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>Private Receipt Sorting with Local AI</title>
      <dc:creator>Swetabh Tripathy</dc:creator>
      <pubDate>Sun, 18 Oct 2026 06:40:30 +0000</pubDate>
      <link>https://dev.to/tripathyswetabh/private-receipt-sorting-with-local-ai-4jgn</link>
      <guid>https://dev.to/tripathyswetabh/private-receipt-sorting-with-local-ai-4jgn</guid>
      <description>&lt;h1&gt;
  
  
  Pocket Ledger: Private Receipt Sorting with Local AI
&lt;/h1&gt;

&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week2-2026-10-12"&gt;Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Pocket Ledger is a local-first receipt and invoice organizer built for freelancers and small businesses.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Freelancers often waste time manually copying expense details from receipts into spreadsheets. Pocket Ledger helps simplify that workflow by extracting vendor names, dates, amounts, and expense categories from receipt text, allowing users to review the results and export bookkeeping records as CSV.&lt;/p&gt;

&lt;p&gt;The workflow is straightforward:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Paste receipt text or upload a text file.&lt;/li&gt;
&lt;li&gt;Extract and normalize expense information.&lt;/li&gt;
&lt;li&gt;Review the extracted fields and correct mistakes.&lt;/li&gt;
&lt;li&gt;Optionally use a local Ollama model to suggest missing information.&lt;/li&gt;
&lt;li&gt;Export the reviewed records to CSV.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The goal is to reduce repetitive bookkeeping without requiring users to send sensitive financial information to a cloud AI provider.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Run Pocket Ledger locally:&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/Tswetabh/Husltr.git
&lt;span class="nb"&gt;cd &lt;/span&gt;Husltr
pip &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;-r&lt;/span&gt; requirements.txt
ollama pull gemma3:4b
streamlit run app.py
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Ollama is optional. Deterministic parsing and CSV export work without it.&lt;/p&gt;

&lt;p&gt;The application is designed to run locally rather than requiring a hosted AI backend.&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;GitHub repository:&lt;/strong&gt; &lt;a href="https://github.com/Tswetabh/Husltr" rel="noopener noreferrer"&gt;https://github.com/Tswetabh/Husltr&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The repository includes:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;app.py&lt;/code&gt; — Streamlit user interface.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;receipt_sorter.py&lt;/code&gt; — Receipt parsing, normalization, and CSV export.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;ollama_client.py&lt;/code&gt; — Integration with a locally running Ollama model.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;tests/&lt;/code&gt; — Automated tests for parsing and model-fallback behavior.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;requirements.txt&lt;/code&gt; — Python dependencies.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;Pocket Ledger combines Python, Streamlit, pytest, and local AI inference.&lt;/p&gt;

&lt;p&gt;The core parsing and export pipeline uses Python's standard library. For optional AI enrichment, the application uses Ollama with the open-weight &lt;code&gt;gemma3:4b&lt;/code&gt; model by default.&lt;/p&gt;

&lt;p&gt;The model receives receipt text and returns structured JSON suggestions. Those suggestions are normalized before being presented for review. Incomplete records are flagged for human review rather than being silently treated as complete.&lt;/p&gt;

&lt;p&gt;A key design decision was to keep AI optional. If Ollama is unavailable, the application falls back to deterministic extraction, allowing the basic workflow to continue without a running model.&lt;/p&gt;

&lt;p&gt;This gives users a practical balance between predictable processing and AI-assisted enrichment.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;Receipt data can contain sensitive financial information, vendor details, addresses, and purchasing habits. Sending that information to a third-party AI service may be unnecessary for a relatively simple bookkeeping workflow.&lt;/p&gt;

&lt;p&gt;Using an open-weight model locally makes several things possible:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Privacy:&lt;/strong&gt; Receipt text can remain on the user's machine when the application and model run locally.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cost control:&lt;/strong&gt; There are no per-document AI API charges.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Flexibility:&lt;/strong&gt; Users can experiment with compatible local models.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Independence:&lt;/strong&gt; The core workflow does not depend on a single AI provider's API or pricing.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Customization:&lt;/strong&gt; The extraction process can evolve to support different bookkeeping conventions and expense categories.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Local AI is not just an alternative implementation. It helps make privacy and predictable operating costs part of the product itself.&lt;/p&gt;

&lt;p&gt;The application still requires a machine capable of running the selected model. Local inference is not automatically faster or more accurate than every hosted alternative, but it gives users greater control over the trade-offs.&lt;/p&gt;

&lt;h2&gt;
  
  
  My First User
&lt;/h2&gt;

&lt;p&gt;I haven't tested Pocket Ledger with a paying user yet.&lt;/p&gt;

&lt;p&gt;My first validation target is a freelancer who spends time each month cleaning up receipt exports and manually organizing expenses.&lt;/p&gt;

&lt;p&gt;I'd start with redacted sample receipts and ask them to try the workflow. I'd specifically measure:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;How much manual work the application eliminates.&lt;/li&gt;
&lt;li&gt;How often extracted fields need correction.&lt;/li&gt;
&lt;li&gt;Which CSV format their accountant or bookkeeping tool requires.&lt;/li&gt;
&lt;li&gt;Whether local processing and privacy would influence their decision to use or pay for it.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The next step is to validate the workflow with a real user before making stronger claims about time savings or willingness to pay.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>Skywatch: An Offline-First Gemma Stargazing PWA Where Screen Time is the Enemy</title>
      <dc:creator>Swetabh Tripathy</dc:creator>
      <pubDate>Mon, 12 Oct 2026 05:17:44 +0000</pubDate>
      <link>https://dev.to/tripathyswetabh/skywatch-an-offline-first-gemma-stargazing-pwa-where-screen-time-is-the-enemy-jo</link>
      <guid>https://dev.to/tripathyswetabh/skywatch-an-offline-first-gemma-stargazing-pwa-where-screen-time-is-the-enemy-jo</guid>
      <description>&lt;h2&gt;
  
  
  Skywatch: An Offline-First, Voice-First Stargazing PWA That Keeps the Screen the Shortest Part of Your Night
&lt;/h2&gt;

&lt;p&gt;&lt;em&gt;This is an official submission for the &lt;a href="https://dev.to/challenges/hf26"&gt;Hacktoberfest Open-Source AI Challenge: Week 1 &amp;amp; 2 ("Touch Grass")&lt;/a&gt;, targeting the **Overall Winner&lt;/em&gt;* and &lt;strong&gt;Best Use of Gemma&lt;/strong&gt; categories.*&lt;/p&gt;




&lt;h2&gt;
  
  
  🌌 The Problem: Stargazing Apps Keep Your Face Buried in Blue Light
&lt;/h2&gt;

&lt;p&gt;Most stargazing apps are interactive planetarium simulators. They expect you to hold a brightly glowing screen up to the sky, pinch-to-zoom across thousands of catalog items, and continuously interact with the display. &lt;/p&gt;

&lt;p&gt;There's just one problem: &lt;strong&gt;it takes 20 to 30 minutes for human eyes to produce rhodopsin and adapt to dark skies&lt;/strong&gt;. Every time you look down at a mobile UI, you reset your dark adaptation. You went outside to look at the universe, but ended up staring at glass.&lt;/p&gt;

&lt;p&gt;We built &lt;strong&gt;Skywatch&lt;/strong&gt; to invert this paradigm: &lt;strong&gt;open-weight AI that gets people outside and makes the screen the shortest part of the entire experience.&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Plan your night ➔ Find best dark spot ➔ Get spoken cues ➔ Capture &amp;amp; ask ➔ Get a Night Card showing how little screen time you used.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;






&lt;h2&gt;
  
  
  🌿 Why Open Innovation Matters Here
&lt;/h2&gt;

&lt;p&gt;When building for the outdoors, closed-source cloud APIs fundamentally fail:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Zero Backcountry Signal:&lt;/strong&gt; The best stargazing locations (Bortle 1–3 state parks, mountain ridges, nature reserves) have zero cellular connectivity. A cloud LLM API like OpenAI or Claude produces a loading spinner in the dark.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Location Privacy:&lt;/strong&gt; Stargazers frequently observe in isolated, remote clearings at night. Transmitting real-time GPS coordinates to commercial cloud servers is an unnecessary privacy exposure.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Local Tool Orchestration with Gemma:&lt;/strong&gt; By running &lt;strong&gt;Gemma 3&lt;/strong&gt; locally via Ollama (&lt;code&gt;http://localhost:11434&lt;/code&gt;), inference happens completely on-device. Gemma acts as a strict tool-calling orchestrator without fabricating astronomy facts.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Single-Config Model Swapping:&lt;/strong&gt; Stargazers on low-power devices can swap to &lt;code&gt;gemma:2b&lt;/code&gt; or step up to &lt;code&gt;gemma3:4b&lt;/code&gt; with a single command (&lt;code&gt;npm run swap-model -- &amp;lt;model&amp;gt;&lt;/code&gt;), without modifying application code.&lt;/li&gt;
&lt;/ol&gt;




&lt;h2&gt;
  
  
  🏗️ Architecture: Gemma as a Grounded Tool-Calling Planner
&lt;/h2&gt;

&lt;p&gt;Gemma must &lt;strong&gt;never&lt;/strong&gt; invent astronomical phenomena. It orchestrates deterministic offline tools that compute ground truth:&lt;br&gt;
&lt;/p&gt;

&lt;pre data-lang="mermaid"&gt;&lt;code&gt;flowchart TD
    subgraph UI ["Client PWA (Offline-First, Vite + React + Tailwind)"]
        PV[1. Plan View]
        SV[2. Dark Sky Scout]
        WV[3. Watch Session - Red-Light Mode]
        CV[4. Capture &amp;amp; Ask]
        NC[5. Night Card - Screen Seconds]
    end

    subgraph Planner ["Open-Weight Local AI Loop"]
        LLM["Gemma 3 (via local Ollama)"]
        DET[Deterministic Astronomy Fallback]
        ZOD[Zod Schema Validation]
    end

    subgraph Tools ["Offline Computation Tools"]
        T_SKY["get_sky_events()"]
        T_WTH["get_weather()"]
        T_SCOUT["score_locations()"]
        T_CTX["get_user_context()"]
    end

    subgraph Engines ["Offline Astronomy &amp;amp; Physics Engines"]
        ASTRO["astronomy-engine (Ephemeris, Moon Phase, Alt/Az)"]
        SAT["satellite.js SGP4 Propagator (ISS Passes)"]
        METEOR["Meteor Catalog (Perseids, Geminids, Orionids)"]
        SCOUT_ENG["Dark Sky Scout (Bortle Grid + Moon Wash + Clouds)"]
    end

    PV --&amp;gt; LLM
    LLM --&amp;gt; T_SKY &amp;amp; T_WTH &amp;amp; T_SCOUT &amp;amp; T_CTX
    T_SKY --&amp;gt; ASTRO &amp;amp; SAT &amp;amp; METEOR
    T_SCOUT --&amp;gt; SCOUT_ENG
    LLM -.-&amp;gt;|Offline / Timeout| DET
    DET --&amp;gt; ZOD
    LLM --&amp;gt; ZOD
    ZOD --&amp;gt; PV
    PV --&amp;gt; WV
    WV --&amp;gt; NC&lt;/code&gt;&lt;/pre&gt;



&lt;h3&gt;
  
  
  The Tool-Calling Agent Loop
&lt;/h3&gt;

&lt;p&gt;When a user asks for a 30-minute stargazing session:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Model proposes tool calls:&lt;/strong&gt;
&lt;/li&gt;
&lt;/ol&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="w"&gt;   &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"tool_call"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"tool"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"get_sky_events"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"arguments"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"lat"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mf"&gt;40.71&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="nl"&gt;"lon"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="mf"&gt;-74.0&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;App executes offline mathematics:&lt;/strong&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;astronomy-engine&lt;/code&gt; calculates real-time altitudes, azimuths, and magnitudes for Venus, Jupiter, Saturn, Mars, and the Moon.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;satellite.js&lt;/code&gt; propagates the ISS orbit via SGP4 from a cached NORAD Two-Line Element (TLE) set.&lt;/li&gt;
&lt;li&gt;Evaluates active annual meteor shower radiants (Quadrantids, Lyrids, Perseids, Geminids, Orionids).&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Model synthesizes the timed plan:&lt;/strong&gt;

&lt;ul&gt;
&lt;li&gt;Generates a timed hit list staggered across the user's available window.&lt;/li&gt;
&lt;li&gt;Computes a calculated &lt;strong&gt;"Leave By"&lt;/strong&gt; time giving observers a 15-minute window to transit and settle night vision.&lt;/li&gt;
&lt;li&gt;Validates all output against a strict &lt;strong&gt;Zod&lt;/strong&gt; schema (&lt;code&gt;SkyPlanSchema&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;If Ollama is offline or returns invalid JSON, the app instantly and seamlessly falls back to the deterministic astronomy engine without failing.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;/ol&gt;




&lt;h2&gt;
  
  
  🎯 The 7 Core Features Built for "Touch Grass"
&lt;/h2&gt;

&lt;h3&gt;
  
  
  1. PLAN: Configurable Agent Modes
&lt;/h3&gt;

&lt;p&gt;Controlled by a modular &lt;code&gt;modes.json&lt;/code&gt; config:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Beginner Stargazer:&lt;/strong&gt; High-contrast naked-eye landmarks and bright planets with simple cardinal cues.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Family &amp;amp; Kids:&lt;/strong&gt; Engaging storytelling hooks, Moon crater spotting, and ISS satellite flybys.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Quick Look (20 min):&lt;/strong&gt; Laser-focused 20-minute hit list prioritizing the highest-altitude bright targets.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Poor Weather / Urban:&lt;/strong&gt; Filters for high-magnitude objects (Moon, Jupiter, Venus) that punch through light smog or partial cloud breaks.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  2. SCOUT: Dark Sky Scout
&lt;/h3&gt;

&lt;p&gt;Combines an offline light-pollution atlas grid calibrated to the &lt;em&gt;New World Atlas of Artificial Night Sky Brightness (Falchi et al., 2016)&lt;/em&gt;, cached cloud cover, and topocentric Moon sky-wash into a physics-backed &lt;strong&gt;0–100 score&lt;/strong&gt; with a crisp one-line reason (e.g. &lt;em&gt;"Pristine Bortle 2 dark sky with crystal atmospheric clarity; Milky Way structure visible"&lt;/em&gt;).&lt;/p&gt;

&lt;h3&gt;
  
  
  3. WATCH: Ultra-Dark Red-Light Mode
&lt;/h3&gt;

&lt;p&gt;Designed for one-hand use in the dark:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;True black background (&lt;code&gt;#050202&lt;/code&gt;) with soft OLED crimson accents (&lt;code&gt;#ef4444&lt;/code&gt;, &lt;code&gt;#fca5a5&lt;/code&gt;) to preserve rhodopsin dark adaptation.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;One Big Tap Target:&lt;/strong&gt; Start session with a single thumb tap.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Stays Completely Quiet:&lt;/strong&gt; Speaks through TTS only when a target is approaching (&lt;em&gt;"ISS visible in two minutes. Look northwest."&lt;/em&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;DeviceOrientation Compass:&lt;/strong&gt; Uses the device magnetometer to provide relative directional cues (&lt;em&gt;"Look slightly to your right, about halfway up. Jupiter is visible."&lt;/em&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Voice Commands:&lt;/strong&gt; Supports &lt;em&gt;"What's that?"&lt;/em&gt;, &lt;em&gt;"What's next?"&lt;/em&gt;, &lt;em&gt;"How long?"&lt;/em&gt;, and &lt;em&gt;"Tell me about this"&lt;/em&gt;, backed by large one-tap dark voice cards for windy nights.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  4. CAPTURE &amp;amp; ASK: Grounded Ephemeris Q&amp;amp;A
&lt;/h3&gt;

&lt;p&gt;Users can snap an outdoor photo and ask questions (&lt;em&gt;"Did I capture a meteor streak or a satellite?"&lt;/em&gt;). Skywatch grounds answers strictly against the computed sky math for that exact second and coordinate. &lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Scientific Honesty:&lt;/strong&gt; The local Gemma model runs in text-only mode and does not perform optical feature detection. The app honestly discloses this to the user and evaluates physical plausibility (e.g. matching against active ISS passes or meteor shower radiant geometry) rather than hallucinating false confirmation.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h3&gt;
  
  
  5. KEEP: Night Cards &amp;amp; The Headline Stat
&lt;/h3&gt;

&lt;p&gt;At the end of a session, Skywatch generates an official &lt;strong&gt;Night Card&lt;/strong&gt;. &lt;/p&gt;

&lt;p&gt;&lt;strong&gt;SCREEN-SECONDS is the headline stat.&lt;/strong&gt;&lt;br&gt;&lt;br&gt;
Instead of rewarding endless app engagement, Skywatch measures real foreground screen time using the browser's &lt;code&gt;visibilitychange&lt;/code&gt; API and active display timers. An observer can spend 45 minutes outside under the stars with &lt;strong&gt;only 42 screen-seconds&lt;/strong&gt; logged!&lt;/p&gt;
&lt;h3&gt;
  
  
  6. COMMUNITY: One-Tap Circles (No Feeds, No Chat)
&lt;/h3&gt;

&lt;p&gt;Minimal nearby stargazing circles where users can RSVP with one tap. Backed by a clean mock dataset (&lt;code&gt;communitySessions.json&lt;/code&gt;) with zero algorithmic distraction, zero followers, and zero chat.&lt;/p&gt;
&lt;h3&gt;
  
  
  7. STREAKS: Touching Grass Consistently
&lt;/h3&gt;

&lt;p&gt;Consecutive-night outdoor streaks and cumulative hours under dark skies calculated purely from stored Night Cards in IndexedDB.&lt;/p&gt;


&lt;h2&gt;
  
  
  ✈️ The Proofs: Verification &amp;amp; Integrity
&lt;/h2&gt;
&lt;h3&gt;
  
  
  1. Airplane-Mode Proof (Zero Outbound Leaks)
&lt;/h3&gt;

&lt;p&gt;We built an automated offline verification script (&lt;code&gt;npm run verify-offline&lt;/code&gt;) and unit test suite that locks the network and tests the entire pipeline:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;========================================================================
✈️  SKYWATCH AIRPLANE-MODE OFFLINE VERIFICATION PROOF
Validating zero-network execution across Plan, Scout, Watch &amp;amp; Night Card
========================================================================

Testing Step 1: PLAN (Offline Astronomy Generation)...
Testing Step 2: SCOUT (Offline Dark Sky Grid Ranking)...
Testing Step 3: WATCH (Compass Orientation &amp;amp; Screen Tracking)...
Testing Step 4: KEEP / NIGHT CARD &amp;amp; STREAKS...

📋 OFFLINE VERIFICATION SUMMARY CHECKLIST
┌─────────┬─────────────────────────┬───────────┬────────────────────────────────────────────────────────┐
│ (index) │ Pipeline                │ Status    │ Details                                                │
├─────────┼─────────────────────────┼───────────┼────────────────────────────────────────────────────────┤
│ 0       │ '1. PLAN'               │ '✅ PASS' │ Generated 2 celestial targets without network.         │
│ 1       │ '2. SCOUT'              │ '✅ PASS' │ Ranked dark sky spots with Bortle &amp;amp; Moon math.         │
│ 2       │ '3. WATCH'              │ '✅ PASS' │ Compass relative heading &amp;amp; screen-seconds verified.    │
│ 3       │ '4. KEEP &amp;amp; STREAKS'     │ '✅ PASS' │ Saved Night Card. Headline stat: 42s screen-time.       │
│ 4       │ '5. NETWORK LEAK CHECK' │ '✅ PASS' │ 0 outbound network calls were made across the pipeline.│
└─────────┴─────────────────────────┴───────────┴────────────────────────────────────────────────────────┘
🎉 100% OFFLINE PROOF COMPLETE: Zero network activity.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  2. Model-Swap &amp;amp; Benchmark Proof
&lt;/h3&gt;

&lt;p&gt;Switch models with a single CLI command:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm run swap-model &lt;span class="nt"&gt;--&lt;/span&gt; gemma:2b
npm run swap-model &lt;span class="nt"&gt;--&lt;/span&gt; gemma3:4b
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;And run side-by-side prompt benchmarks:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm run compare-models &lt;span class="nt"&gt;--&lt;/span&gt; gemma3:4b gemma:2b
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Outputs side-by-side latency, target counts, Zod schema validation passes, and plan quality across tactical, beginner, and family prompts.&lt;/p&gt;




&lt;h2&gt;
  
  
  🔗 Links &amp;amp; Code
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;GitHub Repository:&lt;/strong&gt; &lt;a href="https://github.com/Tswetabh/sky-watch" rel="noopener noreferrer"&gt;https://github.com/Tswetabh/sky-watch&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Live Deployment:&lt;/strong&gt; Deployable as a static PWA on Render / Vercel / Netlify (&lt;code&gt;dist/&lt;/code&gt;)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Challenge:&lt;/strong&gt; &lt;a href="https://dev.to/challenges/hf26"&gt;Hacktoberfest Open-Source AI Challenge: Week 1 &amp;amp; 2&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  🛠️ Try It Yourself
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# 1. Clone &amp;amp; install&lt;/span&gt;
git clone https://github.com/Tswetabh/sky-watch.git
&lt;span class="nb"&gt;cd &lt;/span&gt;sky-watch
npm &lt;span class="nb"&gt;install&lt;/span&gt;

&lt;span class="c"&gt;# 2. Pull local Gemma model&lt;/span&gt;
ollama pull gemma3:4b

&lt;span class="c"&gt;# 3. Run test suite &amp;amp; airplane verification&lt;/span&gt;
npm &lt;span class="nb"&gt;test
&lt;/span&gt;npm run verify-offline

&lt;span class="c"&gt;# 4. Start local PWA&lt;/span&gt;
npm run dev
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Open &lt;code&gt;http://localhost:5173&lt;/code&gt; on your mobile browser, tap "Add to Home Screen", put your phone in your pocket, and go look up at the stars. 🌌🔭&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>Kyūmei: Why I Taught a Local Gemma Model to Refuse to Guess on Hardware Faults</title>
      <dc:creator>Swetabh Tripathy</dc:creator>
      <pubDate>Sun, 04 Oct 2026 22:14:51 +0000</pubDate>
      <link>https://dev.to/tripathyswetabh/kyumei-why-i-taught-a-local-gemma-model-to-refuse-to-guess-on-hardware-faults-5ffa</link>
      <guid>https://dev.to/tripathyswetabh/kyumei-why-i-taught-a-local-gemma-model-to-refuse-to-guess-on-hardware-faults-5ffa</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-weekend-2026-10-01"&gt;Hacktoberfest Weekend Challenge: Build for a Friend&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Kyūmei: A Local Gemma 4 Agent That Cites Evidence Before Naming a Cause
&lt;/h2&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;TL;DR&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;What:&lt;/strong&gt; Kyūmei turns a messy hardware complaint, or a HWiNFO64 sensor log, into numbered evidence, hypotheses that must cite that evidence, and troubleshooting steps ordered from safest to most invasive. It runs Gemma 4 E2B (or Gemma 2 9B) locally through Ollama.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The trick:&lt;/strong&gt; Gemma reasons. Plain code validates the JSON schema, drops causes that cite evidence that doesn't exist, and forces a hard "stop and get it repaired" answer for battery swelling and burning-smell cases.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It can say no:&lt;/strong&gt; If there are fewer than two verifiable clues, it returns &lt;code&gt;needs_more_info&lt;/code&gt; and provides 1-click follow-up questions instead of guessing.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Results:&lt;/strong&gt; 4/4 unit tests on the guardrails pass. On 8 benchmark test scenarios, 7 passed schema validation on the first try and 1 recovered cleanly via the automatic 1-turn retry loop.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;For:&lt;/strong&gt; &lt;strong&gt;Rohit&lt;/strong&gt;, whose 3-year-old gaming laptop was screaming at idle and dying mid-game.&lt;/li&gt;
&lt;/ul&gt;
&lt;/blockquote&gt;




&lt;h2&gt;
  
  
  The 2 a.m. problem
&lt;/h2&gt;

&lt;p&gt;My friend Rohit had an Asus TUF gaming laptop that was acting up during our Hacktoberfest hackathon prep:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;"Bro, the bottom feels like a hot frying pan even when I'm just looking at VS Code, the fans sound like a jet taking off at idle, and after 20 minutes it just shuts off with zero warning."&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;He'd already spent an evening googling. He got the usual soup of generic forum advice: update your BIOS, reinstall Windows, update graphics drivers, run an antivirus scan, or the dreaded &lt;em&gt;"your motherboard VRM is dead, buy a new laptop."&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Nobody was completely wrong. Nobody was &lt;em&gt;useful&lt;/em&gt; either. A laundry list of everything that can possibly go wrong isn't a diagnosis. A diagnosis says: &lt;strong&gt;here is what you told me, here is the exact evidence that points to this specific cause, and here is the cheapest, non-invasive check that would prove me wrong.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;So I built that, and named it &lt;strong&gt;Kyūmei (究明)&lt;/strong&gt; — Japanese for &lt;em&gt;"to investigate the truth / thorough examination."&lt;/em&gt;&lt;/p&gt;




&lt;h2&gt;
  
  
  What it does
&lt;/h2&gt;

&lt;p&gt;You describe the problem the way a real human would:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;Laptop gets very hot on the bottom and fan is loud while idle, shutting down abruptly after 20 minutes of light use.&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Kyūmei processes the input and returns a structured, citation-backed diagnostic report:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Evidence&lt;/strong&gt;: Atomic facts extracted from your symptom, each with an identifier (&lt;code&gt;E1&lt;/code&gt;, &lt;code&gt;E2&lt;/code&gt;, &lt;code&gt;E3&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Candidate Causes&lt;/strong&gt;: Hypotheses where every single cause &lt;strong&gt;must&lt;/strong&gt; cite at least one valid evidence ID.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Verification&lt;/strong&gt;: A concrete physical or sensor check to validate or disprove each cause.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ordered Remediation&lt;/strong&gt;: Non-invasive checks first (software/airflow), internal disassembly last.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A refusal to guess&lt;/strong&gt;: If you provide vague inputs like &lt;em&gt;"my laptop died"&lt;/em&gt;, it returns &lt;code&gt;needs_more_info&lt;/code&gt; with interactive follow-up questions.&lt;/li&gt;
&lt;/ol&gt;

&lt;h3&gt;
  
  
  A Real Diagnostic Run
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;INPUT: "Laptop gets very hot on the bottom and fan is loud while idle, shutting down abruptly after 20 minutes of light use."

[E1] "Laptop gets very hot on the bottom"
[E2] "fan is loud while idle"
[E3] "shutting down abruptly after 20 minutes"

[C1] Heatsink exhaust fins clogged with compacted dust or lint blanket.
     Evidence: [E1], [E2]  |  Confidence: HIGH
     Verification: Inspect exhaust vents with a flashlight to verify lint blockage.

[C2] Degraded or dried thermal interface compound (TIM) on CPU/GPU die.
     Evidence: [E1], [E2], [E3]  |  Confidence: HIGH
     Verification: Log HWiNFO64 sensor data; check if delta temp jumps to 95°C+ within 3s.

TROUBLESHOOTING PLAN:
  1. (Low Risk)   Ensure operation on a hard, flat surface to rule out vent smothering.
  2. (Low Risk)   Blow short bursts of compressed air into exhaust vents while powered down.
  3. (Safe)       Inspect Task Manager for rogue background compute / crypto-miners.

STATUS: OK
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It also reads raw telemetry: export a &lt;strong&gt;HWiNFO64 sensor log&lt;/strong&gt; (&lt;code&gt;.csv&lt;/code&gt; or &lt;code&gt;.txt&lt;/code&gt;), drop it in, and Kyūmei automatically extracts CPU package temps, PROCHOT throttling flags, GPU hot spot temps, fan RPMs, NVMe drive wear, and battery degradation before piping them into the reasoning pipeline. The UI features four dedicated tabs: &lt;strong&gt;Hardware Fault Diagnostics&lt;/strong&gt;, &lt;strong&gt;Telemetry Log Analyzer&lt;/strong&gt;, &lt;strong&gt;Game &amp;amp; AI Feasibility Estimator&lt;/strong&gt;, and &lt;strong&gt;Rogue Process / Miner Triage&lt;/strong&gt;.&lt;/p&gt;




&lt;h2&gt;
  
  
  The idea: the model reasons, the code checks
&lt;/h2&gt;

&lt;p&gt;I didn't want "user asks, LLM hallucinates, ship it." Small local models can sound confident while being subtly or dangerously wrong. In hardware diagnostics, that could mean telling someone to puncture a swollen lithium-ion pouch or replace a working motherboard.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;User Symptom / HWiNFO64 Log
            │
            ▼
      FastAPI / CLI
            │
            ▼
  Gemma 4 E2B via Ollama  (guided by Agent Skill &amp;amp; JSON Schema)
            │
            ▼
       Raw JSON ───► JSON Schema Draft 2020-12 ──[Invalid]──► 1 Auto-Retry Loop
            │                                                       │
         [Valid] ◄──────────────────────────────────────────────────┘
            │
            ▼
    Deterministic Guardrails Engine
       ├── Grounding Filter: Drop causes citing phantom evidence IDs
       ├── Battery / Thermal Safety Clamp: Force technician escalation
       └── Remediation Ordering: Safe software checks first, teardown last
            │
            ▼
    Sanitized Diagnostic Report + Interactive Cross-Highlight UI
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Gemma does what language models are best at: parsing messy, human language and identifying semantic relationships. Deterministic Python code does what code is best at: &lt;strong&gt;enforcing non-negotiable rules&lt;/strong&gt;.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Evidence Grounding Invariant:&lt;/strong&gt; If Gemma suggests a cause backed by &lt;code&gt;E4&lt;/code&gt; and &lt;code&gt;E4&lt;/code&gt; was never extracted in the evidence array, the guardrail immediately drops that cause from the output.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Physical Safety Clamp:&lt;/strong&gt; If the symptom mentions battery swelling, a bulging trackpad, popping chassis seams, smoke, or a burning smell, the engine overrides the response status to &lt;code&gt;safety_alert&lt;/code&gt;, sets a high-contrast hazard warning, and locks the plan to &lt;strong&gt;STOP USE IMMEDIATELY &amp;amp; REFER TO A CERTIFIED TECHNICIAN&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ambiguity Triage:&lt;/strong&gt; When someone gives a two-word complaint like &lt;em&gt;"laptop dead"&lt;/em&gt;, the model doesn't guess. The guardrail sets &lt;code&gt;status: "needs_more_info"&lt;/code&gt; and generates targeted, clickable questions (e.g. &lt;em&gt;"Do any charging LEDs light up when plugged in?"&lt;/em&gt;).&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The diagnostic instructions live in a standardized &lt;strong&gt;Agent Skill&lt;/strong&gt; (&lt;code&gt;skill/SKILL.md&lt;/code&gt;) and formal schema (&lt;code&gt;skill/schema.json&lt;/code&gt;), making the skill portable to any agentic runtime.&lt;/p&gt;




&lt;h2&gt;
  
  
  Where the small model broke
&lt;/h2&gt;

&lt;p&gt;During testing with Gemma 4 E2B and Gemma 2 9B on consumer hardware, we encountered three real failure modes:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Citation Hallucinations:&lt;/strong&gt; On longer symptoms, the model would invent plausible causes citing non-existent IDs like &lt;code&gt;["E4", "E99"]&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Schema Inconsistencies:&lt;/strong&gt; The small model occasionally emitted &lt;code&gt;evidence_ids&lt;/code&gt; as a single string instead of a JSON array of strings (&lt;code&gt;"E1"&lt;/code&gt; instead of &lt;code&gt;["E1"]&lt;/code&gt;), or wrapped the payload in conversational markdown commentary.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Premature Certainty on Vague Prompts:&lt;/strong&gt; When given &lt;em&gt;"my screen went black"&lt;/em&gt;, the raw model would sometimes jump straight to &lt;em&gt;"replace the display panel"&lt;/em&gt; instead of asking whether an external monitor works.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Every failure mode became an explicit deterministic check in &lt;code&gt;app/guardrails.py&lt;/code&gt; and &lt;code&gt;app/validator.py&lt;/code&gt;.&lt;/p&gt;

&lt;h3&gt;
  
  
  Test Suite &amp;amp; Benchmark Metrics
&lt;/h3&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Check&lt;/th&gt;
&lt;th&gt;Result&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Guardrail Unit Tests (Schema, missing fields, ungrounded cause pruning, battery safety clamp)&lt;/td&gt;
&lt;td&gt;
&lt;strong&gt;4 / 4 PASS&lt;/strong&gt; (0.42s)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Benchmark Scenarios with Schema-Valid Output on 1st Try&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;7 / 8&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Scenarios Recovered via 1-Turn Auto-Retry Loop&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;1 / 8&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Causes Dropped for Citing Phantom / Missing Evidence&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;3 instances caught &amp;amp; pruned&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Ambiguous Inputs Correctly Triaged to &lt;code&gt;needs_more_info&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;100% (Scenario 04 &amp;amp; vague edge tests)&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Hazardous / Swollen Battery Symptoms Clamped to Safety Alert&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;100% (Scenario 03 &amp;amp; thermal runaways)&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;em&gt;Run &lt;code&gt;python -m pytest tests/&lt;/code&gt; and &lt;code&gt;python tests/run_examples.py&lt;/code&gt; to reproduce.&lt;/em&gt;&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;An honest caveat:&lt;/strong&gt; Grounding proves that every candidate cause has a verified paper trail in the evidence. It does not prove the model picked the &lt;em&gt;ultimate&lt;/em&gt; root cause. That is why every single hypothesis includes a tangible verification step.&lt;/p&gt;
&lt;/blockquote&gt;




&lt;h2&gt;
  
  
  What Rohit said
&lt;/h2&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;"The fact that it highlighted the exact words from my rambling complaint and connected them to the clogged heatsink made so much sense. We blew out the exhaust with compressed air, and the idle temps dropped from 89°C to 48°C. Best part: it didn't tell me to reinstall Windows."&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;




&lt;h2&gt;
  
  
  Why open, and why local
&lt;/h2&gt;

&lt;p&gt;Hardware diagnostics often involve sensitive telemetry: serial numbers, background processes, user directories, and crash dumps. You shouldn't have to upload your device logs to a closed cloud API.&lt;/p&gt;

&lt;p&gt;Because Gemma is open-weight and runs locally through &lt;strong&gt;Ollama&lt;/strong&gt;, the entire reasoning pipeline operates &lt;strong&gt;100% offline with zero data leakage&lt;/strong&gt;. We tested this on an &lt;strong&gt;NVIDIA GeForce RTX 3050 Laptop GPU (6 GB VRAM)&lt;/strong&gt; running on Windows 11 with Python 3.14 — completely offline with Wi-Fi disabled.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;A note on the hosted showcase:&lt;/strong&gt; For users without local Ollama setups, the repository includes an optional cloud hybrid mode powered by OpenRouter (&lt;code&gt;google/gemma-2-9b-it&lt;/code&gt;). The live demo on Embarko uses this hybrid mode so anyone can test it directly from their browser, but local mode remains the first-class privacy standard.&lt;/p&gt;




&lt;h2&gt;
  
  
  How I built it, and how I used AI
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Backend:&lt;/strong&gt; Python 3.10+, FastAPI, Uvicorn, Pydantic v2, &lt;code&gt;jsonschema&lt;/code&gt; (Draft 2020-12).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Inference:&lt;/strong&gt; Ollama (&lt;code&gt;gemma4:e2b&lt;/code&gt;, &lt;code&gt;gemma2:9b&lt;/code&gt;, &lt;code&gt;llama3.1:8b&lt;/code&gt;) with optional OpenRouter cloud fallback.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Frontend:&lt;/strong&gt; Vanilla modern HTML5, CSS3 with glassmorphism and bidirectional hover cross-highlighting, and reactive ES6 JavaScript.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;AI Tooling Disclosure:&lt;/strong&gt; Built using &lt;strong&gt;Google Antigravity IDE (Gemini)&lt;/strong&gt; as an AI pair programmer for scaffolding boilerplate, CSS polish, and documentation structuring. The core diagnostic reasoning axioms, Agent Skill specification, schema invariants, deterministic guardrails, HWiNFO64 parser, and test suite were designed and implemented by the human author. A full file-by-file audit is published in &lt;code&gt;AI_DISCLOSURE.md&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  Limitations
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Kyūmei is a triage and diagnostic assistant, not a replacement for a certified repair technician.&lt;/li&gt;
&lt;li&gt;Physical safety cases (swollen batteries, electrical burning) strictly halt automated troubleshooting.&lt;/li&gt;
&lt;li&gt;Telemetry parsing currently targets HWiNFO64 &lt;code&gt;.txt&lt;/code&gt; and &lt;code&gt;.csv&lt;/code&gt; logs; screenshot OCR is planned for a future release.&lt;/li&gt;
&lt;li&gt;Evaluated on an 8-scenario benchmark suite on consumer RTX 3050 6GB hardware.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  What's next
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Persistent Local Case History:&lt;/strong&gt; Storing past diagnostic sessions locally so follow-up sensor logs track whether a repair actually fixed the thermal delta.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Screenshot / Thermal Camera OCR:&lt;/strong&gt; Ingesting mobile photos of BIOS screens, BSOD error codes, and FLIR thermal camera captures.&lt;/li&gt;
&lt;/ol&gt;




&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Local Setup (100% Offline with Ollama)
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/Tswetabh/Agent-Kyumei.git
&lt;span class="nb"&gt;cd &lt;/span&gt;Agent-Kyumei/kyumei
python &lt;span class="nt"&gt;-m&lt;/span&gt; venv .venv
.venv&lt;span class="se"&gt;\S&lt;/span&gt;cripts&lt;span class="se"&gt;\a&lt;/span&gt;ctivate      &lt;span class="c"&gt;# On Linux/macOS: source .venv/bin/activate&lt;/span&gt;
pip &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;-r&lt;/span&gt; requirements.txt
ollama pull gemma4:e2b
python &lt;span class="nt"&gt;-m&lt;/span&gt; uvicorn app.main:app &lt;span class="nt"&gt;--host&lt;/span&gt; 127.0.0.1 &lt;span class="nt"&gt;--port&lt;/span&gt; 8000
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Source Code:&lt;/strong&gt; &lt;a href="https://github.com/Tswetabh/Agent-Kyumei" rel="noopener noreferrer"&gt;github.com/Tswetabh/Agent-Kyumei&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Live Demo (Embarko):&lt;/strong&gt; &lt;a href="https://agent-kyumei.embarko.app" rel="noopener noreferrer"&gt;agent-kyumei.embarko.app&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Showcase Listing:&lt;/strong&gt; &lt;a href="https://embarko.ai/showcase/app/agent-kyumei" rel="noopener noreferrer"&gt;embarko.ai/showcase/app/agent-kyumei&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cloud Mirror:&lt;/strong&gt; &lt;a href="https://agentkyumei.vercel.app" rel="noopener noreferrer"&gt;agentkyumei.vercel.app&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  Prize categories
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Best Use of Gemma:&lt;/strong&gt; Gemma 4 E2B runs locally via Ollama to handle messy diagnostic natural-language reasoning, while strict deterministic Python code enforces JSON schema validation, evidence-grounding invariants, and physical safety clamping.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;em&gt;Built during Hacktoberfest Hack Day Indore (PyData Indore × MLH).&lt;/em&gt;&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>weekendchallenge</category>
      <category>hf26challenge</category>
    </item>
  </channel>
</rss>
